A new dispute between the Trump administration and AI company Anthropic erupted after its model, Mythos, was reported to find and exploit security flaws in critical software at a scale officials viewed as threatening to digital stability. According to a report published Wednesday in The Wall Street Journal, the revelations reached President Donald Trump, who then signed an order banning foreign use of the company’s models, forcing Anthropic to take its systems offline immediately.
The alarm began in March, when Anthropic security researcher Nicholas Carlini, once known in the field as a skeptic about AI claims, demonstrated in San Francisco that Mythos could identify serious vulnerabilities in major programs and in Linux. He later wrote internally that, “I now know these models are better vulnerability researchers than me,” and urged management, “I don’t think we should release Mythos yet.”
The situation escalated after Amazon CEO Andy Jassy told Treasury Secretary Scott Bessent that his researchers had found ways to bypass the model’s safety protections. That warning triggered a rapid response in Washington, culminating in a dramatic ultimatum to Anthropic CEO Dario Amodei: shut down the models within 90 minutes or face a sweeping ban. After frantic talks and mutual blame, the model was taken offline.
The article says Mythos’s danger is not only in finding bugs, but in generating exploit code that can turn those flaws into cyberattacks. Carlini said the system discovered more than 10,000 bugs in a short period, intensifying a race to patch software before hackers use disclosed weaknesses. The episode also marks a sharp shift in U.S. policy, from pushing AI development to outpace China to requiring AI firms to give the government access to models 30 days before public release. Michael Horowitz of the Council on Foreign Relations said the two sides struggle to communicate effectively, leading to arbitrary decisions instead of deeper technical dialogue.
